Scheduling algorithms for efficient gather operations in distributed heterogeneous systems

Advances in networks and the availability of low-cost PCs and workstations has resulted in PC/workstation clusters being used as a computer environment that replaces parallel computers. These clusters recently comprise different kinds of nodes and are connected over different kinds of networks, such as Ethernet and Myrinet. They are called distributed heterogeneous systems, and efficient collective communication operations for them have been studied. We propose heuristic scheduling algorithms for efficient gather operations in distributed heterogeneous systems, and we present the results of simulations evaluating their performance. These results show that gather operations scheduled by these algorithms are more efficient than sequential gather operations in these systems. They also show that, the efficiency of gather operations scheduled by the algorithms depends on the ratio of the cost that depends on message size in point-to-point communication to the cost that does not depend on message size in point-to-point communication